Improved signal processing and sensors for classifying clutter and unexploded ordnance (UXO) in areas with overlapping signatures are common is needed. This project investigated the effectiveness of signal processing techniques based on wavelets (1) to improve the signal to noise ratio and extract additional information from the signals and (2) as part of a probability-based approach for classification.
Applying probabilistic methods for classification involves several steps. First, a set of descriptors that adequately distinguishes between classes must be identified. Next, the statistical properties of the descriptor set for each class must be determined, usually by measuring the descriptors for samples taken from members of each class and calculating the usual statistical quantities such as the average and standard deviation. Finally, a statistical test must be developed using the descriptor statistical properties to determine class membership of an unknown sample.
This project’s original intent was to perform all three steps for a limited number of samples of UXO and clutter. Unfortunately, because of unexpected difficulty in performing the wavelet filtering portion of the project, the researchers focused only on the identification of the descriptor set. It has been the researchers’ experience that if an adequate descriptor set can be identified (i.e., one that can distinguish between the classes) then the other two steps involved in applying probabilistic classification methods will present little difficulty. By focusing efforts on the key step in the process, the researchers can effectively evaluate the potential for applying probabilistic methods for classification.
The results of this project are as follows:
The researchers recognize that the number of UXO and clutter samples prevents a statistically valid demonstration on the ability of the descriptor set for classification. However, because the limited results are so unambiguous, there is potential for identifying a descriptor set that will accurately classify UXO and clutter.
Based on this project’s results, the researchers believe that both wavelet filtering and application of probabilistic methods for classification have potential for improving the detection and identification of UXO. Potential follow-on work includes: